Abstract

Pixel classification systems rely on a certain set of features that are sufficient to classify a given pixel into a class defined within a database. Unlike brightness and spectral signature features commonly used in remote sensing applications, texture-based features cannot be defined for a single pixel and must be derived from an area or window surrounding that pixel. In this research, all features are derived from binary morphological granulometries. Once generated, these features comprise a database which can be used to classify images. A Gaussian Maximum Likelihood Classifier is trained with this data base for subsequent classification of both dependent and independent data. Several aspects of these texture-base features require investigation in order to determine their ability to distinguish image textures. Three important aspects are addressed in this study; the effects of maximum noise, the optimal size of the localized window, and the minimum number of optimal features required for accurate classification. A statistical approach has been taken to determine the classification accuracy with varying window size, varying number of features, and varying amounts of four types of maximum noise using granulometric features. Analysis of these investigations indicate four main results. First, classification accuracy in the absence of noise is extremely high. Second, for these textures at the spatial resolution of 75 dpi, classification accuracy decreases dramatically below a window size of 11x11 pixels. Third, the number of features needed for high classification accuracy can be reduced to a fairly small number on the order of 6 features. Finally, these features are generally robust in the presence of maximum noise if the type and amount of noise can be accurately estimated.

Library of Congress Subject Headings

Image processing--Digital techniques

Publication Date

5-1-1991

Document Type

Thesis

Department, Program, or Center

Chester F. Carlson Center for Imaging Science (COS)

Advisor

Vaez-Pravani, Mendi

Comments

Note: imported from RIT’s Digital Media Library running on DSpace to RIT Scholar Works. Physical copy available through RIT's The Wallace Library at: TA1632.N494 1991

Campus

RIT – Main Campus

Download

Share

COinS